Abstract

In the field of saliency detection, many graph-based  algorithms heavily depend on the accuracy of the  pre-processed superpixel segmentation, which leads to  significant sacrifice of detail information from the input  image. In this paper, we propose a novel bottom-up  saliency detection approach that takes advantage of both  region-based features and image details. To provide more  accurate saliency estimations, we first optimize the image  boundary selection by the proposed erroneous boundary  removal. By taking the image details and region-based  estimations into account, we then propose the regularized  random walks ranking to formulate pixel-wised saliency  maps from the superpixel-based background and  foreground saliency estimations. Experiment results on  two public datasets indicate the significantly improved  accuracy and robustness of the proposed algorithm in  comparison with 12 state-of-the-art saliency detection  approaches.